Automated in-hive monitoring and advanced data analytics to detect honey bee diseases

In the UK honey bees contribute over £430m per annum to agriculture, with the Western honey bee (Apis mellifera) providing up to 50% of pollinator ecosystem services. Some suggest the honey bee to be the third most important domesticated animal in the UK, with about 300k colonies and over 40k beekeepers. Unfortunately, honey bees have been badly affected by numerous interacting pressures in recent decades, including agricultural intensification, land use change, extreme weather events, and a growing number of pests and diseases.

This project focuses on two of these disease problems: Varroosis and chronic bee paralysis. The honey bee mite (Varroa destructor) arrived in the UK in the early 1990s, and its presence causes the indigenous deformed wing virus to become highly pathogenic, a condition known as Varroosis. This is the most serious cause of honey bee loss worldwide, and control measures include integrated pest management using pyrethroids to control mites, although mite resistance is becoming increasingly problematic. Chronic bee paralysis is a disease of adult bees caused by the chronic bee paralysis virus (CBPV). This disease is more recent, increasing in prevalence since in the UK since 2007. As such, there are currently no evidence-based control measures for this disease.

Beekeepers need to monitor colonies carefully and on a regular basis to detect Varroosis or chronic bee paralysis. However, it is highly probable that the two diseases are present within a colony before symptoms can readily be detected through routine monitoring of mite populations or visible evidence of diseased worker bees. Previous research has demonstrated that the behaviours of affected bees starts to change in relatively subtle ways at early stages of infection, and the behaviour of healthy bees towards their diseased sisters also alters. New, non-invasive hive-monitoring technologies provide the opportunity to detect some of these changes in behaviour, in particular the acoustics and vibration patterns within a colony. Additional data related to hive temperature, humidity, foraging worker flight exits/return counts, as well as local meteorological conditions can also be collected.

We plan to collect in-hive and external meteorological data from both research and commercial apiaries in the UK across a season. We will also undertake regular assessments of the health of the colonies for Varroosis and symptoms of chronic bee paralysis. Our monitoring will produce large data streams from the apiaries monitored. The largest datasets will be the acoustic data from microphones in the colonies, and we will simplify the raw data to a frequency domain via Fast Fourier Transformation (FFT). The FFT data will then be modelled by either time-series autoregression or machine learning approaches, incorporating other in-hive and external monitor streams as 'meta-data'. The two modelling approaches will be compared, to determine both their effectiveness to discriminate between diseased and healthy colonies, and also their ability to detect disease at an early stage.

We will work closely with our industrial partner (Agrisound) to implement the most practical and cost-effective in-field monitoring systems, in terms of energy use, bandwidth for data streaming etc. This will also include the most appropriate data processing pipelines (centralised or decentralised) to increase the practical value of system.

Varroosis and chronic bee paralysis are major diseases of honey bees in the UK. Regular monitoring of mite populations infestation levels within colonies are needed for Varroosis, whilst bees affected by chronic bee paralysis can be missed by beekeepers. It is likely that much more subtle changes in colony activity occurs as a result of these two diseases at an early stage, before obvious visible symptoms become apparent. We propose to utilise new electronic monitoring technologies, and modern data analytics, for the early diagnosis of these two diseases using the following steps.
Remote capturing of multispectral data from many sensors within and outwith the hive using portal non-invasive monitoring devices.
Deploy monitoring devices to 40 colonies with differing disease states, varying from healthy to diseased and capturing time-series data from an entire beekeeping season.
Determine the levels of disease by monitoring using visual inspections and qPCR.
Develop robust data processing pipelines to diagnose disease from monitoring data. Here acoustic data will be pre-processed through Fast Fourier Transform (FFT). Two approaches will be used: a) FFT simplification through multivariate methods such as PCA, followed by autoregressive time-series analysis that incorporates all the other in-hive and external data as covariables. b) creation of spectrograms from the FFTs for hourly slice, used as inputs into a 4-dimensional convolutional network (CNN) with time as the 4th dimension. A parallel ML model will use the in-hive and external monitoring data, and the two combined. Both models will output colony health for the two diseases.
Establish an 'open air hackaton' to determine if any alternative methods improve the detection of bee disease compared to the above methods.
Finally, design disease monitoring hardware, firmware, and software to monitor disease. The outline design of an efficient system will be developed, that optimises energy use, bandwidth, local an